proofbundle 0.5.0__tar.gz → 0.7.0__tar.gz

{proofbundle-0.5.0/src/proofbundle.egg-info → proofbundle-0.7.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: proofbundle
-Version: 0.5.0
+Version: 0.7.0
 Summary: Emit and verify portable cryptographic evidence bundles, offline: Ed25519 + RFC 6962 Merkle + optional SD-JWT.
 Author: Konrad Gruszka
 License: MIT
@@ -55,17 +55,21 @@ signed and anchored in a tamper-evident log — and optionally carries a
 selectively disclosable credential. Pure Python, no server, no daemon, one JSON file.**
 
 [![CI](https://github.com/b7n0de/proofbundle/actions/workflows/ci.yml/badge.svg)](https://github.com/b7n0de/proofbundle/actions/workflows/ci.yml)
-[![PyPI](https://img.shields.io/pypi/v/proofbundle.svg?color=D6248A)](https://pypi.org/project/proofbundle/)
-[![Python](https://img.shields.io/pypi/pyversions/proofbundle.svg?color=D6248A)](https://pypi.org/project/proofbundle/)
+[![PyPI](https://img.shields.io/pypi/v/proofbundle.svg?color=D6248A&cacheSeconds=3600)](https://pypi.org/project/proofbundle/)
+[![Python](https://img.shields.io/pypi/pyversions/proofbundle.svg?color=D6248A&cacheSeconds=3600)](https://pypi.org/project/proofbundle/)
+[![Downloads](https://static.pepy.tech/badge/proofbundle)](https://pepy.tech/project/proofbundle)
 [![License: MIT](https://img.shields.io/badge/license-MIT-D6248A.svg)](LICENSE)
 [![Ruff](https://img.shields.io/endpoint?url=https://raw.githubusercontent.com/astral-sh/ruff/main/assets/badge/v2.json)](https://github.com/astral-sh/ruff)
 [![SLSA build provenance](https://img.shields.io/badge/SLSA-build_provenance-D6248A.svg)](https://slsa.dev)
+[![PyPI attestations](https://img.shields.io/badge/PyPI-attestations_(PEP_740)-D6248A.svg)](https://pypi.org/project/proofbundle/)
+<!-- DOI badge placeholder: Zenodo is linked and archives each release. Add the Zenodo concept-DOI badge
+     here (and the DOI to CITATION.cff) once Zenodo assigns it — it does not exist at build time. -->
 
 </div>
 
 **At a glance:** `proofbundle emit` signs and anchors a payload; `proofbundle
 verify` checks one self-contained `bundle.json` with three offline cryptographic
-checks → `OK` or `FAILED`. No network, no daemon, no own crypto. 62 tests.
+checks → `OK` or `FAILED`. No network, no daemon, no own crypto. 63 tests.
 
 ## Contents
 
@@ -286,13 +290,19 @@ commitments — it does **not** prove the evaluation was well designed or that t
 itself is correct. Those are human judgements; what it removes is the need to simply
 trust the number.
 
-### Since v0.5: framework adapter, in-toto, selective disclosure
+### A verification layer for trustworthy eval logs
 
-- **inspect_ai adapter** (`pip install "proofbundle[inspect]"`) reads a UK AISI
-  [inspect_ai](https://github.com/UKGovernmentBEIS/inspect_ai) eval log via the stable
-  `read_eval_log` API (lazy import; the core stays dependency-free) and maps it to a claim.
-  `proofbundle.adapters.from_lm_eval_results` reads lm-evaluation-harness `results.json`
-  without importing anything.
+The UK AISI inspect_ai team names an open gap ([arXiv:2507.06893](https://arxiv.org/abs/2507.06893)):
+a database of trustworthy evaluation results with proper provenance tracking. proofbundle is the
+missing **signature + selective-disclosure layer** for exactly that — complementary to metadata
+aggregation (Every Eval Ever) and documentation taxonomies (Eval Factsheets), not a competitor.
+See [INTEROP.md](INTEROP.md) for how it maps to OpenSSF Model Signing, CycloneDX ML-BOM, and in-toto.
+
+- **Two framework adapters** — `pip install "proofbundle[inspect]"` reads a UK AISI
+  [inspect_ai](https://github.com/UKGovernmentBEIS/inspect_ai) eval log via the stable `read_eval_log`
+  API (lazy import). `proofbundle.adapters.from_lm_eval_results` reads a real EleutherAI
+  [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) `results_*.json` (the
+  genuine `acc,none` filter-suffix format) and captures run provenance — no framework import either way.
 - **in-toto Statement v1** — `proofbundle.intoto.to_intoto_statement(claim, root_b64=…)`
   emits the receipt as an in-toto statement with a self-hosted predicate type. The subject
   digest is an *honest salted commitment* under a custom key, never `sha256` (see
@@ -303,6 +313,9 @@ trust the number.
   bundle payload is the source of truth; the SD-JWT is a derived, bundle-bound view, verified
   by proofbundle's own verifier **and** the `sd-jwt-python` reference.
 
+Every release ships **PEP 740 attestations** (Trusted Publishing) + an SLSA build-provenance
+attestation — see [SECURITY.md](SECURITY.md).
+
 ## Roadmap
 
 - **v0.1** — the offline verifier plus a real example bundle.
@@ -310,8 +323,11 @@ trust the number.
 - **v0.3** — external RFC 6962 conformance vectors + real Sigstore Rekor interop.
 - **v0.4** — the eval-receipt emitter (`emit_eval_receipt` / `proofbundle emit-eval`),
   salted commitments, issuer binding.
-- **v0.5 (current release)** — inspect_ai adapter (stable API), in-toto Statement v1 view,
-  and SD-JWT **issuance** per RFC 9901 (selective disclosure of the exact score).
+- **v0.5** — inspect_ai adapter (stable API), in-toto Statement v1 view, SD-JWT **issuance** (RFC 9901).
+- **v0.6** — a second eval adapter (lm-evaluation-harness, real format + provenance), INTEROP.md,
+  CITATION.cff, PEP 740 attestations documented.
+- **v0.7 (current release)** — citability polish: ORCID in CITATION.cff, a Zenodo DOI placeholder
+  (assigned on release), and a draft in-toto ML-eval predicate proposal.
 - **Deferred** (explicitly not yet built) — SD-JWT VC conformance + `vct` metadata,
   Key-Binding JWT, status lists / revocation, an official in-toto PR, DSSE / a full in-toto client.
 

{proofbundle-0.5.0 → proofbundle-0.7.0}/README.md

@@ -12,17 +12,21 @@ signed and anchored in a tamper-evident log — and optionally carries a
 selectively disclosable credential. Pure Python, no server, no daemon, one JSON file.**
 
 [![CI](https://github.com/b7n0de/proofbundle/actions/workflows/ci.yml/badge.svg)](https://github.com/b7n0de/proofbundle/actions/workflows/ci.yml)
-[![PyPI](https://img.shields.io/pypi/v/proofbundle.svg?color=D6248A)](https://pypi.org/project/proofbundle/)
-[![Python](https://img.shields.io/pypi/pyversions/proofbundle.svg?color=D6248A)](https://pypi.org/project/proofbundle/)
+[![PyPI](https://img.shields.io/pypi/v/proofbundle.svg?color=D6248A&cacheSeconds=3600)](https://pypi.org/project/proofbundle/)
+[![Python](https://img.shields.io/pypi/pyversions/proofbundle.svg?color=D6248A&cacheSeconds=3600)](https://pypi.org/project/proofbundle/)
+[![Downloads](https://static.pepy.tech/badge/proofbundle)](https://pepy.tech/project/proofbundle)
 [![License: MIT](https://img.shields.io/badge/license-MIT-D6248A.svg)](LICENSE)
 [![Ruff](https://img.shields.io/endpoint?url=https://raw.githubusercontent.com/astral-sh/ruff/main/assets/badge/v2.json)](https://github.com/astral-sh/ruff)
 [![SLSA build provenance](https://img.shields.io/badge/SLSA-build_provenance-D6248A.svg)](https://slsa.dev)
+[![PyPI attestations](https://img.shields.io/badge/PyPI-attestations_(PEP_740)-D6248A.svg)](https://pypi.org/project/proofbundle/)
+<!-- DOI badge placeholder: Zenodo is linked and archives each release. Add the Zenodo concept-DOI badge
+     here (and the DOI to CITATION.cff) once Zenodo assigns it — it does not exist at build time. -->
 
 </div>
 
 **At a glance:** `proofbundle emit` signs and anchors a payload; `proofbundle
 verify` checks one self-contained `bundle.json` with three offline cryptographic
-checks → `OK` or `FAILED`. No network, no daemon, no own crypto. 62 tests.
+checks → `OK` or `FAILED`. No network, no daemon, no own crypto. 63 tests.
 
 ## Contents
 
@@ -243,13 +247,19 @@ commitments — it does **not** prove the evaluation was well designed or that t
 itself is correct. Those are human judgements; what it removes is the need to simply
 trust the number.
 
-### Since v0.5: framework adapter, in-toto, selective disclosure
+### A verification layer for trustworthy eval logs
 
-- **inspect_ai adapter** (`pip install "proofbundle[inspect]"`) reads a UK AISI
-  [inspect_ai](https://github.com/UKGovernmentBEIS/inspect_ai) eval log via the stable
-  `read_eval_log` API (lazy import; the core stays dependency-free) and maps it to a claim.
-  `proofbundle.adapters.from_lm_eval_results` reads lm-evaluation-harness `results.json`
-  without importing anything.
+The UK AISI inspect_ai team names an open gap ([arXiv:2507.06893](https://arxiv.org/abs/2507.06893)):
+a database of trustworthy evaluation results with proper provenance tracking. proofbundle is the
+missing **signature + selective-disclosure layer** for exactly that — complementary to metadata
+aggregation (Every Eval Ever) and documentation taxonomies (Eval Factsheets), not a competitor.
+See [INTEROP.md](INTEROP.md) for how it maps to OpenSSF Model Signing, CycloneDX ML-BOM, and in-toto.
+
+- **Two framework adapters** — `pip install "proofbundle[inspect]"` reads a UK AISI
+  [inspect_ai](https://github.com/UKGovernmentBEIS/inspect_ai) eval log via the stable `read_eval_log`
+  API (lazy import). `proofbundle.adapters.from_lm_eval_results` reads a real EleutherAI
+  [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) `results_*.json` (the
+  genuine `acc,none` filter-suffix format) and captures run provenance — no framework import either way.
 - **in-toto Statement v1** — `proofbundle.intoto.to_intoto_statement(claim, root_b64=…)`
   emits the receipt as an in-toto statement with a self-hosted predicate type. The subject
   digest is an *honest salted commitment* under a custom key, never `sha256` (see
@@ -260,6 +270,9 @@ trust the number.
   bundle payload is the source of truth; the SD-JWT is a derived, bundle-bound view, verified
   by proofbundle's own verifier **and** the `sd-jwt-python` reference.
 
+Every release ships **PEP 740 attestations** (Trusted Publishing) + an SLSA build-provenance
+attestation — see [SECURITY.md](SECURITY.md).
+
 ## Roadmap
 
 - **v0.1** — the offline verifier plus a real example bundle.
@@ -267,8 +280,11 @@ trust the number.
 - **v0.3** — external RFC 6962 conformance vectors + real Sigstore Rekor interop.
 - **v0.4** — the eval-receipt emitter (`emit_eval_receipt` / `proofbundle emit-eval`),
   salted commitments, issuer binding.
-- **v0.5 (current release)** — inspect_ai adapter (stable API), in-toto Statement v1 view,
-  and SD-JWT **issuance** per RFC 9901 (selective disclosure of the exact score).
+- **v0.5** — inspect_ai adapter (stable API), in-toto Statement v1 view, SD-JWT **issuance** (RFC 9901).
+- **v0.6** — a second eval adapter (lm-evaluation-harness, real format + provenance), INTEROP.md,
+  CITATION.cff, PEP 740 attestations documented.
+- **v0.7 (current release)** — citability polish: ORCID in CITATION.cff, a Zenodo DOI placeholder
+  (assigned on release), and a draft in-toto ML-eval predicate proposal.
 - **Deferred** (explicitly not yet built) — SD-JWT VC conformance + `vct` metadata,
   Key-Binding JWT, status lists / revocation, an official in-toto PR, DSSE / a full in-toto client.
 

{proofbundle-0.5.0 → proofbundle-0.7.0}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "proofbundle"
-version = "0.5.0"
+version = "0.7.0"
 description = "Emit and verify portable cryptographic evidence bundles, offline: Ed25519 + RFC 6962 Merkle + optional SD-JWT."
 readme = "README.md"
 requires-python = ">=3.9"

{proofbundle-0.5.0 → proofbundle-0.7.0}/src/proofbundle/__init__.py

@@ -13,7 +13,7 @@ from .emit import emit_bundle, generate_signer
 from .errors import Check, ProofBundleError, VerificationResult
 from .merkle import verify_consistency, verify_inclusion
 
-__version__ = "0.5.0"
+__version__ = "0.7.0"
 
 __all__ = [
     "__version__",

proofbundle-0.7.0/src/proofbundle/adapters/lm_eval.py

@@ -0,0 +1,76 @@
+"""Adapter for EleutherAI lm-evaluation-harness results_*.json (file-based, NO lm_eval import).
+
+Parses the exported result JSON only — no runtime dependency on lm_eval, no runner rebuild.
+
+Real 0.4.x format (validated against a genuine harness run, see tests/fixtures/lm_eval_arc_easy_real.json):
+the metric keys carry a *filter suffix*, e.g. `"acc,none"`, and the standard error is a **sibling** key
+`"acc_stderr,none"` (not nested). So a caller asking for metric `"acc"` is matched against `"acc,none"`
+(or `"acc,<filter>"`). Provenance (git_hash, harness/task version, n-shot) is copied into the receipt's
+optional `provenance` field so a verifier can trace exactly which run produced it.
+"""
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Optional
+
+from ..evalclaim import build_eval_claim
+
+
+def _find_metric(res: dict, metric: str):
+    """Return (value, stderr, matched_key) for `metric`, handling the `metric,<filter>` suffix format.
+
+    Prefers an exact `metric` key, then `metric,none`, then any `metric,<filter>`. The stderr sibling is
+    `metric_stderr,<same filter>`."""
+    if metric in res:                       # bare key (older/simple exports)
+        stderr = res.get(f"{metric}_stderr")
+        return res[metric], stderr, metric
+    if f"{metric},none" in res:
+        return res[f"{metric},none"], res.get(f"{metric}_stderr,none"), f"{metric},none"
+    for key in res:                         # any filter, e.g. metric,custom-filter
+        if key == metric or (key.startswith(f"{metric},") and not key.startswith(f"{metric}_stderr")):
+            flt = key.split(",", 1)[1] if "," in key else "none"
+            return res[key], res.get(f"{metric}_stderr,{flt}"), key
+    return None, None, None
+
+
+def from_lm_eval_results(path, task: str, metric: str, *, comparator: str, threshold: str,
+                         timestamp: str, model_salt: Optional[bytes] = None,
+                         dataset_salt: Optional[bytes] = None):
+    """Read an lm-evaluation-harness results_*.json and build an eval claim for `task`/`metric`.
+
+    `metric` is the bare name (e.g. "acc"); the real key may be "acc,none". The score is read as a STRING
+    to avoid float canonicalization issues. Returns (claim, salts).
+    """
+    data = json.loads(Path(path).read_text(encoding="utf-8"))
+    res = data.get("results", {}).get(task)
+    if res is None:
+        raise ValueError(f"task not found in results: {task!r}")
+    value, stderr, matched = _find_metric(res, metric)
+    if value is None:
+        raise ValueError(f"metric {metric!r} not found in results[{task!r}] "
+                         f"(available: {sorted(k for k in res if ',' in k)})")
+    score = value if isinstance(value, str) else repr(value)
+
+    n_samples = data.get("n-samples", {}).get(task, {})
+    n = int(n_samples.get("effective") or n_samples.get("original") or res.get("sample_len") or 0)
+    cfg = data.get("config", {})
+    model_id = str(cfg.get("model_name") or cfg.get("model") or "unknown")
+    if cfg.get("model_args"):
+        model_id = f"{model_id}::{cfg['model_args']}"   # include args so the commitment pins the exact model
+
+    provenance = {"harness": "lm-evaluation-harness", "matched_metric_key": matched}
+    if data.get("git_hash"):
+        provenance["git_hash"] = str(data["git_hash"])
+    if data.get("versions", {}).get(task) is not None:
+        provenance["task_version"] = str(data["versions"][task])
+    if data.get("n-shot", {}).get(task) is not None:
+        provenance["n_shot"] = str(data["n-shot"][task])
+    if stderr is not None:
+        provenance["stderr"] = repr(stderr) if not isinstance(stderr, str) else stderr
+
+    return build_eval_claim(
+        suite=task, suite_version=str(data.get("versions", {}).get(task, "lm-eval")),
+        metric=metric, comparator=comparator, threshold=threshold, score=str(score), n=n,
+        model_id=model_id, dataset_id=task, issuer="", timestamp=timestamp,
+        provenance=provenance, model_salt=model_salt, dataset_salt=dataset_salt)

{proofbundle-0.5.0 → proofbundle-0.7.0}/src/proofbundle/evalclaim.py

@@ -37,7 +37,7 @@ _MAX_SAFE_INT = 2 ** 53 - 1
 # The exact key set of an eval claim; decode/validate reject anything else.
 _REQUIRED = {"schema", "suite", "suite_version", "metric", "comparator", "threshold",
              "passed", "n", "model_id_commit", "dataset_id_commit", "commit_alg", "issuer", "timestamp"}
-_OPTIONAL = {"context_binding", "ci95", "multiple_testing", "prereg_sha256"}
+_OPTIONAL = {"context_binding", "ci95", "multiple_testing", "prereg_sha256", "provenance"}
 
 __all__ = [
     "EVAL_CLAIM_SCHEMA", "COMMIT_ALG", "canonicalize", "build_eval_claim",
@@ -126,7 +126,7 @@ def build_eval_claim(*, suite: str, suite_version: str, metric: str, comparator:
                      threshold: str, score: str, n: int, model_id: str, dataset_id: str,
                      issuer: str, timestamp: str, context_binding: Optional[str] = None,
                      ci95: Optional[Sequence[str]] = None, multiple_testing: Optional[str] = None,
-                     prereg_sha256: Optional[str] = None,
+                     prereg_sha256: Optional[str] = None, provenance: Optional[dict] = None,
                      model_salt: Optional[bytes] = None, dataset_salt: Optional[bytes] = None):
     """Build a valid eval claim from raw values. Computes `passed` ITSELF from the comparator
     (never trusts the caller), creates salted commitments, and returns (claim, salts) with the
@@ -163,6 +163,8 @@ def build_eval_claim(*, suite: str, suite_version: str, metric: str, comparator:
         claim["multiple_testing"] = multiple_testing
     if prereg_sha256 is not None:
         claim["prereg_sha256"] = prereg_sha256
+    if provenance is not None:
+        claim["provenance"] = provenance
     _reject_non_jcs(claim)
     return claim, {"model_salt": m_salt, "dataset_salt": d_salt}
 

{proofbundle-0.5.0 → proofbundle-0.7.0/src/proofbundle.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: proofbundle
-Version: 0.5.0
+Version: 0.7.0
 Summary: Emit and verify portable cryptographic evidence bundles, offline: Ed25519 + RFC 6962 Merkle + optional SD-JWT.
 Author: Konrad Gruszka
 License: MIT
@@ -55,17 +55,21 @@ signed and anchored in a tamper-evident log — and optionally carries a
 selectively disclosable credential. Pure Python, no server, no daemon, one JSON file.**
 
 [![CI](https://github.com/b7n0de/proofbundle/actions/workflows/ci.yml/badge.svg)](https://github.com/b7n0de/proofbundle/actions/workflows/ci.yml)
-[![PyPI](https://img.shields.io/pypi/v/proofbundle.svg?color=D6248A)](https://pypi.org/project/proofbundle/)
-[![Python](https://img.shields.io/pypi/pyversions/proofbundle.svg?color=D6248A)](https://pypi.org/project/proofbundle/)
+[![PyPI](https://img.shields.io/pypi/v/proofbundle.svg?color=D6248A&cacheSeconds=3600)](https://pypi.org/project/proofbundle/)
+[![Python](https://img.shields.io/pypi/pyversions/proofbundle.svg?color=D6248A&cacheSeconds=3600)](https://pypi.org/project/proofbundle/)
+[![Downloads](https://static.pepy.tech/badge/proofbundle)](https://pepy.tech/project/proofbundle)
 [![License: MIT](https://img.shields.io/badge/license-MIT-D6248A.svg)](LICENSE)
 [![Ruff](https://img.shields.io/endpoint?url=https://raw.githubusercontent.com/astral-sh/ruff/main/assets/badge/v2.json)](https://github.com/astral-sh/ruff)
 [![SLSA build provenance](https://img.shields.io/badge/SLSA-build_provenance-D6248A.svg)](https://slsa.dev)
+[![PyPI attestations](https://img.shields.io/badge/PyPI-attestations_(PEP_740)-D6248A.svg)](https://pypi.org/project/proofbundle/)
+<!-- DOI badge placeholder: Zenodo is linked and archives each release. Add the Zenodo concept-DOI badge
+     here (and the DOI to CITATION.cff) once Zenodo assigns it — it does not exist at build time. -->
 
 </div>
 
 **At a glance:** `proofbundle emit` signs and anchors a payload; `proofbundle
 verify` checks one self-contained `bundle.json` with three offline cryptographic
-checks → `OK` or `FAILED`. No network, no daemon, no own crypto. 62 tests.
+checks → `OK` or `FAILED`. No network, no daemon, no own crypto. 63 tests.
 
 ## Contents
 
@@ -286,13 +290,19 @@ commitments — it does **not** prove the evaluation was well designed or that t
 itself is correct. Those are human judgements; what it removes is the need to simply
 trust the number.
 
-### Since v0.5: framework adapter, in-toto, selective disclosure
+### A verification layer for trustworthy eval logs
 
-- **inspect_ai adapter** (`pip install "proofbundle[inspect]"`) reads a UK AISI
-  [inspect_ai](https://github.com/UKGovernmentBEIS/inspect_ai) eval log via the stable
-  `read_eval_log` API (lazy import; the core stays dependency-free) and maps it to a claim.
-  `proofbundle.adapters.from_lm_eval_results` reads lm-evaluation-harness `results.json`
-  without importing anything.
+The UK AISI inspect_ai team names an open gap ([arXiv:2507.06893](https://arxiv.org/abs/2507.06893)):
+a database of trustworthy evaluation results with proper provenance tracking. proofbundle is the
+missing **signature + selective-disclosure layer** for exactly that — complementary to metadata
+aggregation (Every Eval Ever) and documentation taxonomies (Eval Factsheets), not a competitor.
+See [INTEROP.md](INTEROP.md) for how it maps to OpenSSF Model Signing, CycloneDX ML-BOM, and in-toto.
+
+- **Two framework adapters** — `pip install "proofbundle[inspect]"` reads a UK AISI
+  [inspect_ai](https://github.com/UKGovernmentBEIS/inspect_ai) eval log via the stable `read_eval_log`
+  API (lazy import). `proofbundle.adapters.from_lm_eval_results` reads a real EleutherAI
+  [lm-evaluation-harness](https://github.com/EleutherAI/lm-evaluation-harness) `results_*.json` (the
+  genuine `acc,none` filter-suffix format) and captures run provenance — no framework import either way.
 - **in-toto Statement v1** — `proofbundle.intoto.to_intoto_statement(claim, root_b64=…)`
   emits the receipt as an in-toto statement with a self-hosted predicate type. The subject
   digest is an *honest salted commitment* under a custom key, never `sha256` (see
@@ -303,6 +313,9 @@ trust the number.
   bundle payload is the source of truth; the SD-JWT is a derived, bundle-bound view, verified
   by proofbundle's own verifier **and** the `sd-jwt-python` reference.
 
+Every release ships **PEP 740 attestations** (Trusted Publishing) + an SLSA build-provenance
+attestation — see [SECURITY.md](SECURITY.md).
+
 ## Roadmap
 
 - **v0.1** — the offline verifier plus a real example bundle.
@@ -310,8 +323,11 @@ trust the number.
 - **v0.3** — external RFC 6962 conformance vectors + real Sigstore Rekor interop.
 - **v0.4** — the eval-receipt emitter (`emit_eval_receipt` / `proofbundle emit-eval`),
   salted commitments, issuer binding.
-- **v0.5 (current release)** — inspect_ai adapter (stable API), in-toto Statement v1 view,
-  and SD-JWT **issuance** per RFC 9901 (selective disclosure of the exact score).
+- **v0.5** — inspect_ai adapter (stable API), in-toto Statement v1 view, SD-JWT **issuance** (RFC 9901).
+- **v0.6** — a second eval adapter (lm-evaluation-harness, real format + provenance), INTEROP.md,
+  CITATION.cff, PEP 740 attestations documented.
+- **v0.7 (current release)** — citability polish: ORCID in CITATION.cff, a Zenodo DOI placeholder
+  (assigned on release), and a draft in-toto ML-eval predicate proposal.
 - **Deferred** (explicitly not yet built) — SD-JWT VC conformance + `vct` metadata,
   Key-Binding JWT, status lists / revocation, an official in-toto PR, DSSE / a full in-toto client.
 

{proofbundle-0.5.0 → proofbundle-0.7.0}/tests/test_adapters.py

@@ -9,15 +9,23 @@ TS = "2026-07-01T12:00:00Z"
 
 
 class TestAdapters(unittest.TestCase):
-    def test_lm_eval(self):
-        claim, salts = from_lm_eval_results(FX / "lm_eval_results.json", "hellaswag", "acc",
-                                            comparator=">=", threshold="0.70", timestamp=TS,
+    def test_lm_eval_real_acc_none_format(self):
+        # REAL lm-evaluation-harness 0.4.12 export: metric key is "acc,none", stderr sibling "acc_stderr,none".
+        claim, salts = from_lm_eval_results(FX / "lm_eval_arc_easy_real.json", "arc_easy", "acc",
+                                            comparator=">=", threshold="0.30", timestamp=TS,
                                             model_salt=b"0" * 16, dataset_salt=b"1" * 16)
-        self.assertEqual(claim["suite"], "hellaswag")
-        self.assertEqual(claim["threshold"], "0.70")
-        self.assertTrue(claim["passed"])              # 0.7534 >= 0.70
-        self.assertNotIn("acme/model-x", str(claim))  # id only as salted commitment
-        self.assertEqual(claim["n"], 10042)
+        self.assertEqual(claim["suite"], "arc_easy")
+        self.assertTrue(claim["passed"])                       # acc 0.5 >= 0.30
+        self.assertEqual(claim["provenance"]["matched_metric_key"], "acc,none")  # suffix handled
+        self.assertIn("git_hash", claim["provenance"])         # provenance captured
+        self.assertEqual(claim["provenance"]["n_shot"], "0")
+        self.assertIn("stderr", claim["provenance"])           # sibling stderr, not nested
+
+    def test_lm_eval_missing_metric_lists_available(self):
+        with self.assertRaises(ValueError):
+            from_lm_eval_results(FX / "lm_eval_arc_easy_real.json", "arc_easy", "nonexistent",
+                                 comparator=">=", threshold="0.5", timestamp=TS,
+                                 model_salt=b"0" * 16, dataset_salt=b"1" * 16)
 
     def test_inspect_ai_stable_api(self):
         # Real .eval log fixture, read via the stable inspect_ai.log.read_eval_log API (proofbundle[inspect]).

proofbundle-0.5.0/src/proofbundle/adapters/lm_eval.py

@@ -1,32 +0,0 @@
-"""Adapter for EleutherAI lm-evaluation-harness results.json (file-based, no framework import)."""
-from __future__ import annotations
-
-import json
-from pathlib import Path
-from typing import Optional
-
-from ..evalclaim import build_eval_claim
-
-
-def from_lm_eval_results(path, task: str, metric: str, *, comparator: str, threshold: str,
-                         timestamp: str, model_salt: Optional[bytes] = None,
-                         dataset_salt: Optional[bytes] = None):
-    """Read an lm-evaluation-harness results.json and build an eval claim for `task`/`metric`.
-
-    Expects the standard shape: {"results": {task: {metric: <number>, ...}, ...},
-    "n-samples": {task: {"effective": n}}, "config"/"model_name": ...}. The score is read as a
-    STRING to avoid float canonicalization issues. Returns (claim, salts).
-    """
-    data = json.loads(Path(path).read_text(encoding="utf-8"))
-    res = data.get("results", {}).get(task)
-    if res is None or metric not in res:
-        raise ValueError(f"task/metric not found in results: {task}/{metric}")
-    score = repr(res[metric]) if not isinstance(res[metric], str) else res[metric]
-    n = int(data.get("n-samples", {}).get(task, {}).get("effective")
-            or data.get("n-samples", {}).get(task, {}).get("original") or 0)
-    model_id = str(data.get("model_name") or data.get("config", {}).get("model") or "unknown")
-    return build_eval_claim(
-        suite=task, suite_version=str(data.get("config", {}).get("model_source", "lm-eval")),
-        metric=metric, comparator=comparator, threshold=threshold, score=str(score), n=n,
-        model_id=model_id, dataset_id=task, issuer="", timestamp=timestamp,
-        model_salt=model_salt, dataset_salt=dataset_salt)